Beijing
What you’ll do
- Develop learning-based algorithms to reconstruct structured road vector data using mass crowdsourced vehicle perception records, REM and multi-modal sensor inputs.
- Model road geometry, semantic features, lane connections and global road topology through spatial reasoning and graph networks.
- Combine deep learning, graph modeling and computational geometry to tackle complex urban scene challenges.
- Build scalable automated pipelines for crowdsourced data aggregation, model validation, map optimization and incremental map updates.
- Continuously optimize model accuracy, robustness and generalization under occlusion, variable illumination and unmarked roads.
- Write standardized, maintainable and testable production code with Python/C++, participate in code review and drive team technical iteration.
What we expect from you
- Master or Ph.D. in Computer Science, Electronic Engineering, Robotics or related majors.
- 3+ years’ algorithm development experience in computer vision, spatial modeling, trajectory mining or robotics.
- Solid programming and algorithm capabilities with Python or C/C++; proficient in at least one deep learning framework (PyTorch / TensorFlow preferred).
- Hands-on experience delivering production-level deep learning or visual perception systems.
- Able to independently research ambiguous technical bottlenecks and deliver practical engineering solutions.
- Fluent oral and written communication in both Mandarin and English, excellent team player.
Nice-to-have
- Familiar with topological learning networks: MapTR, VAD, LaneGAP, TopoNet, as well as image stitching and vectorization algorithms.
- Experience with mass vehicle trajectory aggregation and crowdsourced perception data processing.
- Basic exposure to GIS, computational geometry, SLAM or ADAS lightweight vector map development.
- In-depth understanding of CNN, GNN, Transformer, object detection, semantic segmentation and generative AI.
- Proven track record of migrating academic research algorithms to mass-production pipelines.
